In the eyes of many, the birth of ChatGPT marks a great revolution in the field of AI.
At the same time, this AI chatbot, which is popular all over the world, has also kicked off the competition in the field of large models for many AI companies around the world. Among them, Big Tech company Meta is determined to win this race.
(Source: Meta)
It is understood that in recent years, the company has been working hard to roll out generative AI products to billions of users to challenge OpenAI, the creator of ChatGPT, for its leading position in related technologies, which includes transforming computing infrastructure, integrating research and product teams, and more.
In addition, its virtual assistant, Meta AI, launched in September 2023, has also been built into the search box of Instagram, Facebook, WhatsApp, Messenger, and other software, making it easy for users to get answers to their target questions using this AI tool.
However, if Meta wants its AI assistant to truly compete with ChatGPT, it must have the same or even better underlying model as OpenAI. In this regard, Meta has created the Llama series of models.
It is reported that on April 18, 2024 local time, the company launched its latest product, Llama 3. According to reports, the version is divided into the following two models, one is the Llama 3 8B with 8 billion parameters, and the other is the Llama 3 70B with 70 billion parameters.
In terms of model performance, Llama 3 not only outperforms similar models developed by other companies, such as Mistral 7B and Gemma 7B, in many of the most popular core benchmarks today, but also performs very well in tasks such as inference, generating code and instructions, etc.
(Source: Meta)
"Meta equipped Llama 3 with new computer coding capabilities and provided it with images and text in training. Chris Cox, Meta's chief product officer, told the media.
However, it is worth noting that the model can only output text at the moment, and future versions will provide more advanced coding and inference capabilities, such as longer context windows.
According to a blog post published by Meta, the company also plans to release a version with multilingual and multimodal capabilities in the near future, which means that Llama 3 is expected to generate text and images.
"The ultimate goal [of Llama 3] is to help users get out of their troubles and make their lives easier, whether it's interacting with businesses, writing things, or planning trips," Cox told the media. ”
In fact, while Meta now offers Llama 3 as a download for developers to use for free, the company prefers to define the model as "open" rather than "open source."
The reason for this is that even though Meta allows developers to use Llama 3 for academic research or commercial applications, it doesn't come with any strings attached, as users might think.
In other words, Meta prohibits developers from using the model to train other generative models, and requires companies with "more than 700 million monthly active users" to apply for a special license from Meta at its sole discretion.
Meanwhile, Meta is training the Llama 3 model with more than 400 billion parameters, and it could surpass the world's best closed AI models, including OpenAI's GPT-4 and Google's Gemma, according to Yann LeCun, Meta's chief AI scientist.
Currently, the two Llama 3 models mentioned above have been integrated into Meta AI and can be accessed through the independent website of "Meta.ai".
In response, Meta CEO Mark Zuckerberg posted on Instagram: "With this new model, we believe Meta AI is now the smartest AI assistant you can freely use." ”
Specifically, Meta AI can help users with tasks like recommending restaurants, planning trips, polishing emails, and more. Even if you use Meta AI's "Imagine" feature, you can generate very high-definition images faster.
According to reports, the image begins to appear the moment the user types and changes "every few letters entered".
It's worth mentioning that Meta AI has so far only been launched in the United States, but is rolling out an English version to countries such as Australia, Canada, Singapore, and more.
"It seems to me that a lot of people today don't think about Meta AI when they think about the main AI assistants that people use. Zuckerberg told the media, "But I think now is the moment when we start to really introduce it to a lot of people, and I expect it to be a pretty important product." ”
In fact, although Meta has regained the top spot in the open source rankings with Llama 3, this is not the final result at a time when the competition for AI models is so fierce.
After all, there have been public reports that OpenAI is preparing to develop GPT-5, which could help the company outperform other competitors once again. According to Zuckerberg, Meta is already planning to put the development of Llama 4 and Llama 5 on the agenda.
As he told the media, "Our goal is not to compete with open source models, but to compete with all the AI in the world and become the world's leading AI." ”
Resources:
https://www.businessinsider.com/meta-releases-latest-ai-model-llama-3-2024-4
https://about.f.com/News/2024/04/Mehta-I-Assistant-Built-with-Lama-3/
https://llama.meta.com/llama3/
hattapus://ai.meta.com/blog/meta-lama-3/
https://techcrunch.com/2024/04/18/meta-releases-llama-3-claims-its-among-the-best-open-models-available/?guccounter=1&guce_referrer=aHR0cHM6Ly93d3cuZ29vZ2xlLmNvbS8&guce_referrer_sig=AQAAAEE9-OCbYiDOGweIH3W0zBvUcAjFhdi71UcFtdZB_Dl8_-xEkbWq8-pbesL72kILjjtJtnApB6Td6Ql6Sw5arQ2RQtOkj8XvTGfz3ul4-jQDBn63N1lEmBMWNeNSwzcRMJESpv-NHW-hVL969aeHB_5iqGk-Ysz-B1hK2CU8P0Td
https://w.devaraj.kama/2024/4/18/24133808/Mehta-I-Assistant-Lama-3-Satzipt-Makeup-Rival
https://www.cnbc.com/2024/04/18/meta-ai-assistant-comes-to-whatsapp-instagram-facebook-and-messenger.html
https://wu.instagram.com/reel/K56Jugschmis/
https://sg.news.yahoo.com/meta-releases-llama-3-claims-160028613.html?guccounter=1&guce_referrer=aHR0cHM6Ly93d3cuZ29vZ2xlLmNvbS8&guce_referrer_sig=AQAAAByrCjkMUkhkTD5w_0CcrEZmH9-s0sI8NEaLOCmCns7E6Kis1H0vnePskhKX7IoDTII75OcFvpkjHBLfhMCLruhqusDLihB4IAsLQmFqBNU9sMjQb-1MftT_laMHStrDIj4AkZeLti_Jz-wtTf8LD03VoA3XQsUfe99fZIXpTnjF
Operation/Typesetting: He Chenlong